- Learn
- Machine Learning
- Recommender Systems and Ranking
ML-M12
Recommender Systems and Ranking
How products decide what to show you next: the user-item matrix, collaborative filtering and matrix factorization, two-tower retrieval, learning to rank, graph models and cold start.
Every feed, shop and streaming service has far more items than any person can browse. A recommender picks the handful worth showing, and it is often the model that touches the most users in a company.
You will start with the user-item matrix and the two classic ideas: recommend items like the ones a user liked, or items that similar users liked. Then you will factorize that sparse matrix into small user and item vectors, scale retrieval to millions of items with two-tower models and nearest-neighbour search, and order the final list with learning to rank. The module ends with graph-based recommenders and the hard practical problems: new users and items, diversity, and the feedback loops a recommender creates.
Lessons
- ML-12.01 Recommender fundamentals Recommenders predict what a user will like. Beginner 10 min
- ML-12.02 Collaborative filtering and matrix factorization Matrix factorization learns a small vector for every user and every item so that their dot product predicts the interaction. Intermediate 15 min
- ML-12.03 Two-tower retrieval and embedding-based retrieval Large recommenders work in stages: retrieval narrows millions of items to a few hundred candidates, then ranking orders them precisely. Advanced 14 min
- ML-12.04 Learning to rank Learning to rank optimizes the order of results rather than individual scores. Advanced 14 min
- ML-12.05 Graph-based recommendations Users, items and their interactions form a graph. Advanced 10 min
- ML-12.06 Cold start, diversity and feedback loops New users and new items have no history, the cold-start problem, so systems fall back on content features, popularity and onboarding questions. Advanced 10 min